December 2022 Summaries

In a comparison of deep learning hardware for computer vision, Habana Gaudi HPUs have emerged as a cost-effective alternative to the traditionally dominant NVIDIA GPUs. By benchmarking the YOLOv5 model training on the COCO dataset, it was found that the Habana Gaudi1 HPUs achieved a cost efficiency of $0.73 per epoch, outperforming the NVIDIA A100 GPUs, which cost $0.98 per epoch. This represents a 25% cost advantage for the Gaudi1 HPUs, although they are slower in terms of epoch time, indicating their potential for greater economic efficiency in model training. The study, conducted with AWS instances, illustrates the evolving landscape of AI hardware, with the promising Gaudi2 accelerators on the horizon potentially offering further advancements.

In December 2022, Piotr Skalski published a detailed guide on training a YOLOv7 instance segmentation model using a custom dataset. The tutorial highlights the efficiency of the YOLOv7 model, which, despite being less accurate than models like OneFormer, offers significantly faster performance, making it suitable for real-time applications. Skalski outlines the process of setting up a Python environment with GPU access, installing YOLOv7, and inferring with pre-trained models, emphasizing the importance of balancing speed and accuracy in model selection. The guide also details preparing a custom dataset using Roboflow, focusing on the significance of using polygon annotations for precise object shape learning, and discusses the steps for applying data augmentations and exporting the dataset for training. Finally, Skalski explains the evaluation of model performance on unseen data, the deployment of trained models using Roboflow, and the potential applications of instance segmentation in various fields such as structural damage assessment in buildings, highlighting the usefulness of the YOLOv7 model for practical, real-world tasks.

Historically, GPUs have been the preferred choice for computer vision training due to their high performance, but recent advancements in custom-designed CPUs, such as Intel's "Ice Lake" and "Sapphire Rapids," offer promising alternatives. The Intel c6i Ice Lake AWS Instance, powered by 3rd Generation Intel Xeon Scalable processors, provides up to 15% better price performance than previous instances and features a balance of compute resources and cost, making it suitable for various compute-intensive applications. Meanwhile, the Intel Sapphire Rapids R7iz instances, which are built on 4th generation Intel Xeon Scalable processors, are designed for high-performance computing and AI workloads, boasting the highest performance per vCPU among x86-based EC2 instances. In benchmarking experiments comparing these CPU instances to GPU instances for computer vision inference, the Intel c6i.2xlarge instance showcased the best cost-to-performance ratio, though it did not achieve the highest inference speeds. The analysis suggests that while CPU instances like Ice Lake and Sapphire Rapids offer cost-effective alternatives with fewer management complexities, GPU instances remain more suitable for achieving the fastest inference speeds.

James Gallagher's blog post discusses the implementation of computer vision to monitor and analyze customer movement in retail stores, enhancing operational efficiency and customer satisfaction. The approach involves using cameras to track how long customers spend in various store sections like electronics, food, or clothing, which can inform staffing decisions and store layout adjustments. The process detailed in the post includes dividing the store into labeled sections, detecting individuals using object detection technology, and training a model with annotated images to enhance accuracy. By tracking the duration customers spend in specific areas, retail stores can optimize space allocation and reduce congestion, particularly at checkout counters. The tutorial outlines the technical steps for creating a tracking system, including data annotation, model training, and script writing, and suggests potential applications for improving shopper experience and store management. The post encourages retail businesses to utilize computer vision to gain insights into shopper behavior, offering a step-by-step guide to building a custom people-tracking model using Roboflow.

Semi-supervised learning (SSL) is a machine learning technique that combines elements of supervised and unsupervised learning to train models using a small set of labeled data along with a large amount of unlabeled data, addressing the time and cost constraints of manual data labeling. This approach is beneficial for various tasks, including classification, regression, clustering, and anomaly detection, by leveraging assumptions like continuity, cluster, and manifold to find relationships within the dataset. Techniques such as pseudo-labeling, self-training, and label propagation are employed to iteratively refine model predictions by utilizing the confidence of predictions from the partially trained model. Although effective in many applications, SSL requires representative labeled data to perform well, and it is not suitable for all tasks, especially when the labeled data does not represent the full distribution of the dataset.

Inspecting structural damage in buildings and bridges is traditionally a challenging and costly endeavor, but using drones equipped with computer vision models offers a promising solution. By capturing close-range footage, drones can identify structural damage such as cracks and concrete degradation. Using the Roboflow platform, images of damage are collected and labeled, followed by the training of a computer vision model. This model, once trained, can be deployed to detect damage in real-time through a Python script that processes live camera feeds. The effectiveness of the model was demonstrated with high accuracy rates, and it can be expanded to identify other types of structural issues. This innovative method not only streamlines the inspection process but also enhances safety and efficiency by reducing the need for manual evaluations.

Ultralytics and Roboflow have collaborated to create an integration that simplifies the import and export of YOLOv5 models and datasets between the two platforms, enhancing the computer vision pipeline for users. This partnership allows Roboflow users to export datasets directly to Ultralytics HUB for custom model training and then use these models for model-assisted labeling within Roboflow, promoting an active learning pipeline. The integration provides an interoperable solution for computer vision endeavors, enabling seamless transitions between dataset management, model training, and automated labeling, thus streamlining the process of building efficient and effective computer vision applications.

Piotr Skalski explores the use of computer vision to track football players, revisiting a project initially attempted with basketball players, and documents the process in a Roboflow video and blog post. Utilizing a combination of YOLOv5 and ByteTRACK, Skalski aimed to efficiently track players on the field but encountered challenges with pre-trained models detecting irrelevant objects and failing to reliably track the ball. Consequently, he constructed a custom dataset and trained a new model, introducing additional classes like referee and goalkeeper. Despite initial setbacks, the custom-trained model improved detection accuracy, although class imbalance issues persisted. ByteTRACK was employed for its efficient object tracking capabilities without requiring additional training. The project demonstrates the potential for analyzing player movements, tracking ball trajectories, and identifying key field zones, while Skalski encourages others to experiment with the open-source code and dataset available on Roboflow Universe.

Struck-by accidents are a significant hazard on construction sites, and the use of computer vision technology offers a promising solution to reduce these incidents. A project by a Roboflow community member named Troy demonstrates how computer vision can be employed to track the proximity of workers to moving machinery, potentially alerting both workers and drivers to dangerous situations. The project involves using cameras and object detection models to monitor construction sites, identifying machinery and workers, and calculating their distances to prevent accidents. The model processes video frames to detect active equipment and flags workers who are too close, with labels indicating caution. This system could be integrated into construction vehicles to provide real-time alerts, enhancing safety protocols on-site by keeping track of safety incidents and offering a visual representation of site activity. The project highlights the potential of using computer vision to improve safety measures in construction environments, with the possibility of live deployment to monitor and prevent accidents as they happen.

Roboflow's November 2022 update highlights several key advancements and resources in the field of computer vision, with improvements to their Polygon and Smart Polygon labeling tools and new training graphs for object detection. The release of the Roboflow-100 research paper and related datasets marks a significant contribution to object detection benchmarks, complemented by various tutorials, including those available on YouTube. A notable feature is the introduction of over 30 new dataset aggregate pages across various industries, enhancing accessibility to computer vision resources. The company also celebrated milestones with team anniversaries and meet-ups, while expanding job opportunities to further democratize computer vision. The community engagement is evident through 15,000+ new projects on their platform, 32 forum discussions, and 19 blog posts, alongside collaborations with industry giants like Intel and mentions in Amazon Web Services publications, demonstrating Roboflow's growing influence and commitment to advancing machine learning capabilities.

In 2021, Gary Blackwood, a postgraduate student at the University of Glasgow, conducted a study using computer vision to explore correlations between litter amounts in Glasgow and socio-economic factors such as employment rate, income, and education levels. By employing a YOLOv5s object detection model trained on thousands of annotated Google Street View images, Gary aimed to create a reliable data source for analyzing litter distribution across various city areas. His findings, derived from regression analyses, indicated no statistically significant relationship between socio-economic factors and litter rates, although limitations such as hardware constraints and image resolution might have affected the outcomes. Despite these challenges, the study highlighted computer vision's potential in urban policy-making, proposing that the model could be scaled and adapted for various applications, including improved trash collection strategies in Glasgow and potentially other cities.